PERFORMANCE OPTIMIZATION OF LARGE CONTROL SYSTEMS

CASE STUDY ON A CONTINUOUS PULP DIGESTER

K. Halmevaara, H. Hytyniemi
Helsinki University of Technology, Control Engineering Laboratory,
PO Box 5500, 02015 TKK, Espoo, Finland
{kalle.halmevaara, heikki.hyotyniemi}@hut.fi

Abstract: This paper presents an example application of a novel multivariate control

parameter tuning method, called Iterative Regression Tuning (IRT), on a simulator of
continuous pulp digester. IRT is a data based method in which multivariate regression and
iterative optimization methods are utilized in the parameter tuning. In the test case seven
PI controllers and two model based controllers were tuned simultaneously and six userdefined quality measures were set as optimization targets. Encouraging results on the
process performance improvement were obtained and the method proved a clear potential
for optimization of large industrial systems. Copyright 2005 IFAC
Keywords: Multivariable control systems, Large-scale systems, Process control, Pulp
Industry, Multiple-criterion optimization, Iterative methods, Regression analysis,
Complex systems

1. BACKGROUND
The control performance assessment and controller
tuning have been intensive research topics during the
last decades. Undoubtedly, one reason for this is the
fact that as many as about 60 % of the industrial PID
controllers are behaving inefficiently or even
detrimentally, i.e., far from the optimal achievable
control performance, see, e.g., (Harris et al., 1999).
Especially the work initiated by Harris (1989) on the
minimum variance benchmarking has attracted a lot
of interest. Improvements on the original work have
been reported, e.g., in (Huang and Shah, 1999). Also,
methods for detecting oscillating and sluggish
control loops have been published, see, e.g.,
(Thornhill and Hgglund, 1997) and (Hgglund,
1999). In (Tyler and Morari, 1996) a control
performance monitoring method based on likelihood
methods was introduced. Various controller tuning
techniques have been presented by researchers and
many of them have been even applied by the
practicing control engineers, particularly those for
single loop tuning.
One interesting control tuning method was suggested
by Hjalmarsson et al (1994). They introduced the
Iterative Feedback Tuning (IFT) method that takes a
new approach to control performance improvement.

The basic idea is to find the parameters that optimize

the LQG type design criterion. IFT is an example of
the iterative identification in closed loop methods
that are used for controller redesign. These methods
and closed loop identification in general are
reviewed, e.g., in (Landau, 2001).
A more general approach compared to IFT is the
Iterative Regression Tuning (IRT) method, first
described in (Hytyniemi, 2002). A case study, in
which the performance of a power plant simulator
was optimized using the proposed method, was
reported in (Halmevaara and Hytyniemi, 2004).
This paper gives another successful example of
applying IRT to performance optimization. In
Chapter 2 the idea of the IRT method is presented.
Chapter 3 introduces the model of the pulping
process, Chapter 4 presents results of the case study
and in Chapter 5 the results are discussed more
thoroughly and some conclusions are drawn.
2. IRT METHOD
A dynamical MIMO system (multiple inputs,
multiple outputs) with input and output signals, u and
y, respectively, is presented in Fig. 1. The parameters
define how the output signals depend on the inputs.

In general, consists of any continuously

enumerable parameters, such as, controller
parameters, setpoint values or process parameters. In
this paper only control parameters are considered,
however. The performance of the system can be
characterized by means of quality measures, q, that
are calculated from the input and output signals. For
instance, robustness and accuracy are important
concepts in describing the performance of a control
system. Such concepts can be measured, e.g., with
the variance and the setpoint tracking ability of
controlled variables, respectively.
In statistical sense, the quality measures values, q,
depend on the system parameters, . For example, if
all parameters affecting the output are properly
chosen, the system performance will be satisfactory
on the average, regardless of the input signals that
are introduced to the system. Viewed on a lower
abstraction level the system is dynamic determining
the dependency between u and y, whereas the upper
level connection between and q can be considered
static. Although the connection between and q
involves stochastic uncertainties, q being a function
of more or less random realizations of signals u and
y, this dependency can be modelled statistically. This
model makes it possible to optimize the system
performance with respect to its parameters.
Because no physical dependency between and q is
assumed to be known, the modeling is based on
observed data. It turns out that, under certain
assumptions regarding the data distribution, linear
model is optimal for this purpose. First, if the data
can be assumed unimodal, meaning it comes from a
single multivariate Gaussian distribution, the
dependency between and q is linear in the
maximum likelihood sense (see, e.g., Pindyck and
Rubinfeld, 1991). The Gaussianity assumption can
be motivated by the central limit theorem stating that
if a number of independent variables are added
together the resulting distribution approximates
Gaussian, no matter what the original distribution of
the variables was. However, it must be noted that the
Gaussianity assumption holds only locally and
therefore the linear models cannot be applied
successfully over the whole parameter space but only
locally. Not to lose the Gaussianity, the quality
measures are assumed to be smooth functions of the
parameters. The validity of the above assumptions in
practice will be discussed in more details below
within the results of the case study.
As was mentioned above, the system under
consideration was assumed to be a MIMO system.
Consequently, it is reasonable to assume that there
exists a large amount of parameters affecting the
outputs of the system, and similarly multiple quality
measures are required to describe the desirable
performance of the overall system. Modeling the
dependencies of parameters and quality measures
with a linear model has many advantages. First, the

linear model is easy to compute and, second, the

numerical problems typically plaguing estimation of
high dimensional linear models can be avoided
effectively by using multivariate statistical methods,
such as Principal Component Analysis (PCA), Partial
Least Squares (PLS) and Canonical Correlation
Analysis (CCA). These methods are presented more
thoroughly, e.g., in (Hytyniemi 2001).
The required data sets are generated with Markov
Chain Monte Carlo (MCMC) simulations, in which
the values of are randomly varied (locally) and the
consequent values of q are calculated. (Notice, that a
simulator of the process is used here for the data
generation.) If the system involves n parameters and
the performance is assessed with m quality measures,
the vector representation of a data sample is
1
q1
= # , q = # ,

q
n
m

(1)

and the data set of k samples can be expressed

T (1)
qT (1)

= # , Q= # .
T ( k )
qT ( k )

(2)

The multivariate statistical methods assume the data

to be appropriately preprocessed. This means that,
the data has to be locally centered and scaled to unit
variance before modelling. Based on data, a linear
model F that describes the dependency between
and q,

q = F T ,

(3)

can be estimated. The same mapping for the data set

of k samples in the matrix form can be presented as
Q = F .

(4)

System

Evaluation

q
Fig. 1. A higher-level statistical model between
parameters, , and quality measures, q, describes
the relevant properties of the underlying
dynamic MIMO system, having inputs u and
outputs y. (Hytyniemi, 2002)

the length of the parameter update step. The notation

F(K) emphasizes the fact that the matrix F is
estimated over and over again, in every global
iteration step. Also the values of and w can vary
during the optimization procedure.

Fig. 2. The dependency of the parameters 1 and 2

and the quality measure q is modelled based on
data in four successive global iteration steps
(Halmevaara and Hytyniemi, 2004).
The m quality measures can be combined into a
scalar valued optimization cost function, J, e.g., by
calculating a weighted sum,
m

J = wT q = wi qi ,

(5)

i =1

where w is an m 1 weighting vector. Alternatively,

various multiobjective optimization methods are also
available, see, e.g., (Miettinen, 1999). As was
mentioned above, the estimated model, F, is valid
only locally, and therefore the optimal solution,

* = arg min J ( q ( ) ) ,

(6)

cannot be found with a single calculation and one has

to accept the approach of taking short update steps
towards this optimum. The easiest way to find the
local optimum (here, a minimum) is the Gradient
descent algorithm. Other applicable methods are,
e.g., Conjugate gradient method, Newton-Rhapson
and Scoring algorithms. The gradient, which
indicates the direction of the maximal growth of the
cost function, J, is obtained by substituting (3) into
(5) and by differentiating the equation resulting in
dJ
d
=
( wT F T ) = Fw .
d d

(7)

To minimize the value of J the parameters are

updated to negative gradient direction according to
( K + 1) = ( K )

dJ
( K ) = ( K ) F ( K ) w . (8)
d

Above, refers to prevailing parameter values (the

centre of the parameter variation in a local iteration
process), K is the global iteration step index and is

In summary, the IRT method can be condensed as

follows (see Fig. 2): The procedure consists of K
global iteration steps. Through these steps the values
of the parameters are gradually tuned towards their
optimal values. Each step consists of a local iteration
process, i.e., k simulations are run with slightly
varied parameter values and the consequent quality
measures are calculated. A local linear model is
estimated from the data if the Gaussianity
assumption is not violated. The subsequent
parameter update is based on the calculation of the
gradient of the cost criterion. Global iteration steps
are taken until the performance of the system meets
its objectives or as long as significant improvements
on the system performance can be observed.
One of the most important advantages of the IRT
method is that it ignores the structure of the
underlying control system, i.e., no matter whether a
multivariable controller or a multi loop SISO control
system is considered. In both cases, the control
structure is optimized as a whole including all
interdependencies. Optimal performance for large
and complex control systems cannot be achieved by
tuning controllers one by one. Further, the possibility
to use different quality measure definitions makes
the method flexible and expressive.
3. PULP DIGESTER CASE STUDY
A model of a Finnish pulp mill was applied in the
test case. The process model was constructed with
the APROS software that is a professional simulation
environment for modeling combustion and nuclear
power plants, and pulp and paper mills (see Anon.
2004). APROS provides large libraries of process
and automation components, which can be combined
into rigorous models of industrial plants. APROS has
been used successfully, e.g., in various training
simulator and process analysis projects.
In the following, some terminology related to pulp
production is first explained. Then, the model of the
pulp mill is introduced, after which the tuning targets
of the case study are formulated.

3.1. Terminology
The kappa number is a measure of lignin content in
the cooked pulp. It is measured with the amount of
potassium permanganate consumption in an acid
pulp dilution. The remaining lignin content of the
pulp has a strong effect on the colour of the produced
fibres. It also affects the strength properties of pulp,
which makes it an important controlled variable.

Fig. 3. The model of the pulp mill: The chip feed

screw conveyor (left), the impregnation vessel
(in the middle), and the pulp digester (right).
The kappa number is controlled by adjusting the H
factor. H factor is an experimental combination of
the cooking time t and the temperature T. One H
factor unit denotes the effect of one hour cooking in
100 C. The H factor is defined as follows
t

H = e(

43,2 16115 T )

dt .

(9)

level controllers in the impregnation and the digester

vessels were oscillating heavily due to their
inappropriate tuning. Due to unsatisfactory level
control, also the washing coefficient control failed to
meet its targets. The two model-based controllers
were behaving even more detrimentally. The other
one predicts the kappa number of the produced pulp
based on the digester top temperature measurement.
There is a five hour time delay in the cooking
process which makes this prediction essential for the
process control. The other model calculates the
setpoint value for the H factor based on the measured
temperature profile of the digester, the amount of
applied alkali and the kappa number of the produced
pulp. Both models were giving strongly biased
predictions, causing naturally serious problems. Due
to the improper controller tuning the process was
producing continuously out of specification pulp.

3.3. Optimization targets

Cost function of the optimization was formulated
according to equation (5) with m = 6 and wi = 1, i =
16. All quality measures were defined with the
same mathematical expression,

After cooking the pulp is washed in several

successive washing stages before bleaching. The
cooking chemicals are recycled, and thus the
washing losses are tried to minimize. The washing
starts at the bottom of the digester with a counter
current washing. The efficiency of the washing is
measured with the washing coefficient that is defined
as ratio of used washing liquor to the amount of
washed pulp. To ensure uniform pulp properties
constant washing coefficient is desirable.

3.2. Model and problem description

A general overview of the model is presented in Fig.
3. First, the woodchips and the impregnation liquor
are mixed, after which the mixture is fed into the
impregnation vessel. The flow continues to the top of
the digester, where the mixture is heated with steam
to the cooking temperature. Wood was modelled to
consist of several cellulose, carbohydrate and lignin
components, and the liquor was assumed to contain
sodium hydroxide and sodium sulphide in addition to
the organic compounds dissolved from the
woodchips. Chemical reactions during the cooking
phase were modelled according to (Gustafson et al.
1983). Several circulation streams of used black
liquor to the preceding parts of the process were
modelled. The subsequent washing and bleaching
operations were excluded from the model.
In the case study, the production of pulp in a steady
state operation was considered, i.e., changes of
neither production rate nor quality targets of the pulp
were simulated. In the initial state of the test case,
several control loops were behaving poorly. The

qi =

1 T
yi (t ) - yi , sp (t ) ,
T t =1

(10)

in which i = 16, t is the time series sample index, T

is the length of the simulated time series, yi is the
measured process variable and yi,sp is the setpoint
value for yi (except for i = 6, for which yi is the
predicted H factor and yi,sp is the true H factor
calculated from the digester temperature profiles).
The six considered variables were:

y1: kappa number in the digester blow,

y2: washing coefficient,
y3: liquor level in the digester,
y4: chip level in the digester,
y5: chip level in the impregnation vessel, and
y6: H factor prediction.
The tuning involved seven PI controllers and two
models that were applied in prediction and setpoint
calculation, i.e., n = 20. The PI controllers were
responsible for regulation of the chip and liquor
levels, washing coefficient, production rate, H factor,
and the digester steam chamber temperature.
4. RESULTS
This chapter presents the results of K = 30 global
steps. The fairly great number of global steps is due
to the crude gradient method. Also, since the initial
performance was purposefully set to extremely poor,
long tuning time was understandable. In simulations
k = 50 on the average and T = 8h. Simulations were
run about 25 times faster than real time.

Quality measure 1

of the kappa number has stopped, variance of the

digester liquor level has diminished and the H factor
prediction has improved tremendously.

Quality measure 2

0.31

q1

2
0.3

5.5

0.29
0.28

0.27
4.5

Kappa number

0.26

40

0.25

0.24
3.5

10

20

Quality measure 3

0.028

30

0.23

38
0

10

20

Quality measure 4

0.22

30

36

q4

3
0.026

0.2

0.024

34

0.18

0.022
0.02

Setpoint = 33

0.16

32

0.018

0.14

0.016

0.012

30

0.12

0.014
0

10

20

Quality measure 5

0.0145

30

0.1

10

20

Quality measure 6

1100

q5

30

28

q6
1000

0.014

t [h]

900
0.0135

Deviation from the digester liquor levelsetpoint [cm]

800

700

0.013

600
0.0125

500

400

0.012

300
0.0115

10

20

30

200

0
0

10

20

30

-2

Fig. 4. The quality measure values, qi, i = 16, in

the global optimization steps, K = 130.

-4
-6

Cost function J

t [h]

Deviation of the H factor prediction

400

-400
1

10

20

-800

30

-1200

Fig. 5. The values of the cost function, J, during the

iterative optimization, K = 30.

The success of the optimization can be followed also

from the values of J (see Fig. 5). It can be seen how
the conflicting targets finally start to slow down the
tuning procedure. If the objectives are not met at the
end, one has to reconsider the weighting of qi. It is
the only way to continue the optimization among
Pareto optimal solutions. Examples of the improved
performance are presented in Fig. 6. The fluctuation

t [h]

Fig. 6. Simulation results with initial and tuned

parameters, K = 1 (dotted) and K = 30 (solid).
Quality measure distribution: Global step 18

Principal component 2
Principal
component 2

The IRT method succeeded in improving the process

performance regarding the six quality measures (see
Fig. 4). For example, the kappa number deviation, q1,
diminished from about 6 to 4 units, and the absolute
value of the H factor prediction error, q6, was
reduced from about 1000 to 200 units. For softwood
pulp the kappa number is tried to keep within 2-3
units range from the target. In that sense the control
is not yet satisfactory but the tuning succeeded to
improve its performance notably. The quality
measure values include some stochastic variation and
therefore the trends are not monotonically
descending. Also occasional outliers can be
perceived in the values of the quality measures in the
global iteration steps 15-20. The causes behind these
observations are discussed below.

Quality measure distribution: Global step 25

Principal
2
Principalcomponent
component 2

-1

-2

-3
-5

2
1
0
-1
-2
-3

-4

-3

-2
-1
Principal component 1

Principal component 1

-4
-4

-2

2
4
Principal com ponent 1

Principal component 1

Fig. 7. Distributions of q values projected to the

plane spanned by the two major principal
components, K = 18 (left) and K = 30 (right).
The convenient linear model was justified with the
assumption of Gaussian data. The distributions of the
quality measure data in global steps K = 18 and K =
30 are shown in Fig. 7. Obviously the distribution in
the left figure does not fulfil the Gaussianity
assumption. The same can be seen also from Fig. 8
in which the empirical cumulative distribution
functions (ECDF) of the distributions are compared
to the theoretical CDF of the (0,1) Gaussian

distribution. This comparison is known as the

Kolmogorov-Smirnov test. It is evident that at some
points in the parameter space the true unknown cost
function may be strongly nonlinear, and therefore
Gaussianity testing is recommended.
5. CONCLUSIONS
Grasping the general view of a large and complex
system, like an industrial process and its control
system, is a demanding task. The underlying
interdependencies are difficult to comprehend as the
size of a system increases. Advanced multivariate
statistical methods are required to capture the
emerging higher abstraction level concepts.
The IRT method makes it possible to optimize the
performance of large systems, and helps the domain
area experts to refine their intuition concerning the
process and the goals of its performance. As long as
the distribution of the quality measures stays close to
Gaussian, the IRT method proceeds convincingly.
Therefore, defining q should be done carefully.
Departures from Gaussianity can be due to nonlinear
behaviour of the quality measure or too large local
variation of the parameters.
The application of IRT method to controller tuning
assumes that a rigorous simulation model of the
process is available. It is also assumed that the
available computational power enables one to run
simulations notably faster than in real time. In that
case the IRT method opens up new possibilities:
Different control structures can be tuned to their
maximum performance and compared with each
others before implementation, and, e.g., start-ups of
new process plants can be hastened by finding
reasonable initial values for controller parameters. In
future, implementation of IRT to a semantic mill
model design environment will be studied.
ECDF vs. Normal CDF

1
0.9

ECDF(q1F(x)
) vs. CDF

0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
-4

-3

-2

-1
0
x1
ECDF vs. Normal
CDF

1
0.9

ECDF(q1F(x)
) vs. CDF

0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
-5

-4

-3

-2

-1
x1

Fig. 8. ECDF of q1 vs. theoretical CDF of the (0,1)

Gaussian distribution, K = 18 (top) and K = 30
(below).

REFERENCES
Anon. (2004). APROS - The Advanced Process
Simulation Environment, VTT (Technical
Research Center of Finland), (referred
26.8.2004), www.vtt.fi/tuo/63/apros/index.htm.
Gustafson, R.R., C.A. Sleicher, W.T. McKean and
B.A. Finlayson (1983). Theoretical Model of the
Kraft Pulping Process. Industrial & Engineering
Chemistry Process Design and Development, 22,
87-96.
Hgglund, T. (1999). Automatic detection of
sluggish control loops. Control Engineering
Practice, 7, 1505-1511.
Halmevaara, K. and H. Hytyniemi (2004). Iterative
Simulation
Based
Multivariate
Control
Parameter Tuning Method. To be published in:
Proceedings of the 5th EUROSIM Congress on
Modeling and Simulation. Paris, France.
Harris, T.J. (1989). Assessment of control loop
performance. Canadian Journal of Chemical
Engineering, 67, 856-861.
Harris, T.J., C.T. Seppala and L.D. Desborough
(1999). A review of performance monitoring and
assessment techniques for univariate and
multivariate control systems. Journal of Process
Control, 9, 1-17.
Hjalmarsson, H., S. Gunnarsson and M. Gevers
(1994). A Convergent Iterative Restricted
Complexity Control Design Scheme. In:
Proceedings of the 33rd IEEE Conference on
Decision and Control, 1735-1740. Orlando, FL.
Huang, B., and S.L. Shah (1999). Performance
assessment of control loops, Springer Verlag,
London, UK.
Hytyniemi, H. (2001). Multivariate Regression
Techniques and Tools. Helsinki University of
Technology, Control Engineering Laboratory,
Helsinki, Finland.
Hytyniemi, H. (2002). On Emergent Models and
Optimization of Parameters. In: Proceedings of
42nd Scandinavian Simulation Conference
SIMS02, 45-50. Oulu, Finland.
Landau, I.D. (2001). Identification in closed loop: a
powerful design tool (better design models,
simpler controllers). Control Engineering
Practice, 9, 51-65.
Miettinen, K.M. (1999). Nonlinear Multiobjective
Optimization, Kluwer Academic Publishers,
Norwell, MA.
Pindyck, R.S. and D.L. Rubinfeld (1991).
Econometric Models and Economic Forecasts,
3rd edition, McGraw-Hill, New York.
Thornhill, N.F. and T. Hgglund (1997). Detection
and diagnosis of oscillation in control loops.
Control Engineering Practice, 5, 1343-1354.
Tyler, M.L. and M. Morari (1996). Performance
Monitoring of Control Systems using Likelihood
Methods. Automatica, 32, 1145-1162.